Claude Haiku 5.5 Explained: Anthropic’s Fastest Model Just Got Much Cheaper
On 7 October 2026, Anthropic released Claude Haiku 5.5, the third model in its 5.5 family after Opus 5.5 and Sonnet 5.5. Haiku is the small, fast tier of the Claude family — the model built for work that runs thousands of times a day. This release is unusual because the headline is not just capability: Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for most prompts, which Anthropic says works out to roughly 75% cheaper on average than the Haiku 4.5 it replaces.
This explainer covers what Haiku 5.5 is, exactly how its two-tier pricing works (including the tokenizer catch almost nobody mentions), what Anthropic’s benchmarks actually claim, and who should — and should not — switch to it.
Contents
– What is Claude Haiku 5.5?
– What’s new compared with Haiku 4.5
– Haiku 5.5 pricing, explained properly
– The tokenizer catch: recount before you rebudget
– What the benchmarks say (and who ran them)
– Haiku 5.5 vs Sonnet 5.5 vs GPT-6 Luna
– Who should use Haiku 5.5
– Who should not
– How to access Haiku 5.5
– Frequently asked questions
– The bottom line
What is Claude Haiku 5.5?
Every Claude generation ships in three sizes. Opus is the flagship for the hardest reasoning and coding work, Sonnet is the everyday all-rounder, and Haiku is the smallest and fastest — designed for high-volume, latency-sensitive jobs: summarising documents, classifying and tagging content, extracting data, drafting short replies, and doing the repetitive sub-tasks inside larger AI agents.
Haiku 5.5 keeps that role and upgrades the specs around it:
- Context window: 1 million tokens — up from 200K on Haiku 4.5, and the same headline figure as its bigger siblings.
- Maximum output: 128K tokens per response (with batch jobs supporting up to 300K output tokens in beta, per Anthropic’s model documentation).
- Inputs: text and images. Output: text.
- Knowledge cutoff: June 2026.
- Model ID:
claude-haiku-5-5. - Availability: the Claude Platform (API), Amazon Bedrock, Google Cloud and Microsoft Foundry, live from launch day. Haiku is also one of the models available in the Claude chat apps — it is listed on Anthropic’s plan comparison for every tier, including Free.
The design goal is clear from those numbers: Haiku 5.5 is not a cut-down toy model. It is a full-context, long-output model priced for volume.
What’s new compared with Haiku 4.5
| Haiku 4.5 | Haiku 5.5 | |
|---|---|---|
| Input price (per 1M tokens) | $1.00 | $0.10 (prompts up to 100K) / $0.50 (over 100K) |
| Output price (per 1M tokens) | $5.00 | $0.50 / $2.50 |
| Context window | 200K tokens | 1M tokens |
| Effort setting | No | Yes — first Haiku with one (default: medium) |
| Tokenizer | Previous | New — counts ~30% more tokens for the same text |
| OSWorld 2.1 (offline subset) | 15.7% | 72.4% |
| Terminal-Bench 4.0 | 0.0% | 39.2% |
Prices and benchmark scores from Anthropic’s launch materials, checked 18 October 2026.
Three changes matter most:
1. The price collapse. For prompts up to 100K tokens — which Anthropic says covered about 90% of requests to Haiku 4.5 — the per-token price fell by 90%. Above 100K it fell by 50%. Blended across typical workloads, Anthropic puts the average saving at about 75%.
2. The effort setting. Haiku 5.5 is the first Haiku with an effort parameter: you can dial how long the model thinks before answering, trading a little latency for better answers on harder requests, instead of switching model tiers. Adaptive thinking is on by default and the default effort level is medium.
3. Computer and terminal skills. The benchmark jumps (next section) are concentrated in tasks where the model operates a desktop or a command line — exactly the skills a subagent inside a larger system needs. Anthropic shipped beta computer-use and browser-use support in its Python and TypeScript SDKs alongside the launch.
Haiku 5.5 pricing, explained properly
Haiku 5.5 uses a two-tier price split at 100K prompt tokens:
| Per 1M tokens | Prompts up to 100K | Prompts over 100K | Haiku 4.5 (for reference) |
|---|---|---|---|
| Input | $0.10 | $0.50 | $1.00 |
| Output | $0.50 | $2.50 | $5.00 |
| Cache writes | $0.125 | $0.625 | $1.25 |
| Cache reads | $0.01 | $0.05 | $0.10 |
Two more levers cut real-world bills further:
- Prompt caching is extremely cheap at this tier — a cache read costs one cent per million tokens on short prompts.
- Batch processing takes 50% off input and output for work that doesn’t need an instant answer.
Anthropic also cut Sonnet 5.5 cache reads by 50% at the same launch (from $0.20 to $0.10 per million tokens), which it says makes Sonnet 5.5 roughly 20% cheaper on most agentic workloads. If you run a mixed fleet — Sonnet for the hard steps, Haiku for the volume steps — both halves of that fleet got cheaper on the same day.
The tokenizer catch: recount before you rebudget
Here is the detail that will catch teams out. Haiku 5.5 ships with a new tokenizer that counts the same text as roughly 30% more tokens than Haiku 4.5 did. Anthropic’s migration guidance is explicit: token counts, max_tokens settings, usage figures and cost estimates measured on Haiku 4.5 all need to be redone for claude-haiku-5-5.
That doesn’t erase the saving — a 90% price cut absorbs a 30% token increase easily. But it does mean a workload that used exactly 100K tokens on Haiku 4.5 may now count as ~130K and land in the higher price tier, and a budget spreadsheet copied from last quarter will be wrong in both directions. Recount your real prompts before you promise anyone a number.
One more migration note from Anthropic’s documentation: non-default temperature, top_p or top_k values return a 400 error on this model — if your integration sets sampling parameters explicitly, remove them when you switch the model ID.
What the benchmarks say (and who ran them)
The scores below are Anthropic’s own launch-table figures. Vendor benchmarks are run under conditions the vendor chooses, so read them as “Anthropic’s best documented case,” not as independent lab results — but the pattern across them is consistent and worth understanding.
| Benchmark | Haiku 5.5 | Haiku 4.5 | GPT-6 Luna |
|---|---|---|---|
| OSWorld 2.1, offline subset (operating a desktop) | 72.4% | 15.7% | 48.9% |
| Terminal-Bench 4.0 (command-line tasks) | 39.2% | 0.0% | 16.4% |
| FrontierCode 1.1 Main (coding) | 46.4% | — | 42.4% |
| GDPval-AA v2.1 (office-work rating) | 1620 | — | 1437 |
Source: Anthropic’s Haiku 5.5 announcement table, checked 18 October 2026. Anthropic does not state whether it ran GPT-6 Luna itself or used OpenAI’s published figures.
The pattern: Haiku 5.5’s biggest leads are in operating computers and terminals, its smallest lead is in pure coding, and Sonnet 5.5 still beats it clearly on complex agentic coding (70.6% on Terminal-Bench 4.0 in the same table). In other words, the cheap model got dramatically better at doing tasks, while the mid-tier model remains the one Anthropic itself points at for hard, long-horizon coding agents.
Haiku 5.5 vs Sonnet 5.5 vs GPT-6 Luna
- Against GPT-6 Luna: Luna lists identical short-context rates ($0.10/$0.50), but its higher price tier only starts above 272K input tokens, against Haiku’s 100K threshold. For a single very long prompt, Luna can be cheaper on list price; on Anthropic’s shared benchmarks, Haiku 5.5 scores higher on every row the two share. Price per token is not price per finished task — accuracy and retries decide real cost.
- Against Sonnet 5.5: Sonnet costs $2.00/$10.00 per million tokens — twenty times Haiku’s short-prompt rates — and earns it on complex reasoning, long coding sessions and judgement-heavy work. The sensible architecture for most teams is routing: Haiku for the many small steps, Sonnet for the few hard ones.
- Against Haiku 4.5: there is no remaining reason to start new work on 4.5. The successor is cheaper at every tier, has five times the context window, and adds the effort control 4.5 never had.
Who should use Haiku 5.5
- High-volume pipelines — summarisation, classification, extraction, tagging, moderation pre-checks — where per-request cost was the blocker.
- Agent builders who need a cheap subagent model for compaction, summarising and tool-running steps inside a larger system.
- Computer-use and browser-use prototypes, where its benchmark profile is strongest relative to price.
- Chat products that need fast first responses with a genuine 1M-token memory of the conversation.
- Anyone whose Haiku 4.5 bill was mostly short prompts — that is the workload the 90% cut was designed for.
Who should not
- Hard, long-horizon coding agents — Anthropic’s own guidance points to Sonnet 5.5 and Opus 5.5; Haiku’s Terminal-Bench score is a fraction of Sonnet’s.
- Workloads dominated by very long prompts — above 100K tokens the price advantage halves, and Luna’s later threshold can win on list price.
- Teams who can’t re-test — the new tokenizer changes counts everywhere; switching the model ID without re-running your evaluations and budgets is how surprise bills happen.
How to access Haiku 5.5
- API / Claude Platform: set the model to
claude-haiku-5-5. It’s also available through Amazon Bedrock, Google Cloud and Microsoft Foundry. - Claude chat apps: Haiku is one of the models listed for Claude plans, including the Free plan — casual users meet it as the fast option inside claude.ai and the mobile apps.
- If you’re migrating from Haiku 4.5: recount tokens on real traffic, remove explicit sampling parameters, re-run your eval set, then rebudget using the two-tier table above.
Frequently asked questions
When was Claude Haiku 5.5 released?
On 7 October 2026, as the third model in Anthropic’s 5.5 family, after Opus 5.5 and Sonnet 5.5 earlier in the fall.
How much does Claude Haiku 5.5 cost?
$0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100K tokens; $0.50 and $2.50 above that. Cache reads start at $0.01 per million, and batch processing takes a further 50% off. Anthropic says the average workload runs about 75% cheaper than on Haiku 4.5.
What is Haiku 5.5’s context window?
1 million tokens — five times Haiku 4.5’s 200K — with up to 128K tokens of output per response.
Is Haiku 5.5 better than Sonnet 5.5?
No. It’s much cheaper and faster, and it’s the right tool for high-volume, well-scoped tasks. Sonnet 5.5 remains clearly stronger on complex agentic coding and hard reasoning — in Anthropic’s own table it scores 70.6% on Terminal-Bench 4.0 against Haiku 5.5’s 39.2%.
Can I use Haiku 5.5 for free?
Haiku is listed among the models on Claude’s Free plan in the chat apps, subject to the free plan’s usage limits. API use is pay-as-you-go at the rates above; there is no free API tier in the pricing checked for this article.
What does the effort setting do?
It controls how long the model thinks before answering. Haiku 5.5 is the first Haiku to offer it; the default is medium, with adaptive thinking on. Lower effort is cheaper and faster; higher effort suits harder requests.
The bottom line
Haiku 5.5 is a price-and-volume release with a real capability jump attached: a 1M-token context window, the family’s first effort dial on the small model, desktop and terminal skills that leapfrog its predecessor — and a price cut deep enough (90% on the prompts most people actually send) to change which projects are viable at all.
Your next step: if you run anything on Haiku 4.5 today, recount a day of real traffic with the new tokenizer, switch the model ID in a test environment, and compare cost per completed task — not per token — against your current bill. If you’re choosing a first cheap model for a new agent project, Haiku 5.5 and GPT-6 Luna are the two to benchmark, on your own prompts, this month.
Prices and benchmark figures in this article were checked against Anthropic’s launch announcement and model documentation on 18 October 2026 and can change; Anthropic’s documentation is the source of truth.